MiniMax launches H3 Max with first-frame and last-frame control

MiniMax announced yesterday the release of H3 Max, an expanded version of the H3 video model co-developed with fal.ai. The new model adds first-frame and last-frame guidance for both text-to-video and image-to-video generation and is intended to shorten wait times compared with the base version.

Pricing per second and low resolutions

The model produces clips 5 to 15 seconds long in two quality tiers: 480p at $0.05 per second and 768p at $0.08 per second. Six supported aspect ratios range from ultra-wide 21:9 to vertical 9:16, targeting formats used on Reels, TikTok and YouTube Shorts. There is currently no option for 1080p or 4K.

median latency of 32.8 seconds (32.79 seconds at P50)

End-to-end latency stands at 32.79 seconds at the P50 percentile for the fastest provider. This figure measures the time from request submission to receipt of the final file, not just raw inference time. In practice, a 10-second clip at 768p incurs roughly 3.3 seconds of waiting per second of video, an improvement over the previous generation but still far from real-time.

provider availability and automatic routing

According to platform data, the primary provider’s uptime is 100%, while the overall average uptime is 90.85%. When an upstream provider returns an error, the system can automatically route the request to another healthy provider, provided the request filters allow it. Per-provider availability metrics are exposed via the Endpoints API, enabling production applications to implement custom fallback logic.

what’s missing

The announcement does not include visual quality metrics such as FVD, CLIPSIM, or human evaluation, nor a direct comparison with the baseline H3 model or details on additional training data that differentiate the versions. The 15-second clip limit remains, requiring manual stitching for longer videos. For developers building on fal.ai, the per-second pricing, keyframe control, and API-driven provider routing make H3 Max a pragmatic, though not groundbreaking, option that is usable today.